--- title: Universal Computing Research colorFrom: blue colorTo: indigo sdk: static pinned: false license: apache-2.0 emoji: 🌍 --- Background image # Universal Computing Research **Universal Computing Research** is an independent AI research organization focused on efficient, compact, and architecture-driven deep learning. We build small language models, parameter-efficient neural layers, custom tokenizers, and research artifacts that test how far useful intelligence can be pushed under strict compute, memory, and parameter budgets. ## Research direction Our work is centered on a simple question: > How much capability can be recovered through better architecture, tokenization, data curricula, and parameterization, without relying only on scale? Current focus areas: - Small language models - Parameter-efficient architectures - Random projection layers - Custom tokenization pipelines - Arithmetic and algorithmic reasoning ## Released models ### Atom3.4m A 3.41M parameter decoder-only language model trained from scratch for studying compact architectures, curricula, and small-model benchmarking. - Grouped-query attention - RoPE positional embeddings - RMSNorm - Gated SiLU feed-forward layers - Custom 4,096-token byte-level BPE tokenizer - Approximately 5B training tokens [View Atom3.4m](https://huggingface.co/UniversalComputingResearch/Atom3.4m) ### Atom2.7m A 2.74M parameter causal language model with an arithmetic-aware tokenizer and digit-structure features. - Custom byte-level BPE tokenizer - Atomic digit and operator handling - Least-significant-digit-first numeric representation - Place and role embeddings for integer arithmetic - Strong ArithMark-2.0 performance for its size [View Atom2.7m](https://huggingface.co/UniversalComputingResearch/Atom2.7m) ## Research ### Parametrized Random Projection We study **Parametrized Random Projection** layers as lightweight replacements for dense linear layers. The core idea is to separate fixed feature mixing from learnable adaptation: a non-trainable random projection performs the mixing, while small learnable element-wise parameters modulate the input and output. This reduces trainable parameter count from quadratic to linear scale while preserving much of the utility of dense projections. [Read the paper](https://arxiv.org/abs/2512.13480) ## Open source Our models and research artifacts are released to support reproducible, open, and practical AI research.